Determination of urinary stone composition using computed tomography using artificial intelligence

Urinary stone composition analysis is essential for guiding treatment and preventive strategies, but conventional methods such as Fourier-transform infrared spectroscopy (FTIR) are costly and limited in availability. Artificial intelligence (AI), particularly convolutional neural networks (CNNs), offers potential as a non-invasive alternative using computed tomography (CT) imaging. The aim of the study was to evaluate the performance of a CNN model in predicting urinary stone composition from non-contrast CT (NCCT) images.

A retrospective analysis was conducted using an institutional database of patients with urinary stones surgically removed between 2020 and 2024. Stone composition was confirmed via FTIR. A total of 1,487 NCCT scan images were used for CNN development and internal validation, with 105 scan images reserved for validation. Transfer learning with VGG16 architecture was applied, incorporating data augmentation and fine-tuning strategies. Performance metrics included accuracy, area under the ROC curve (AUC), and Cohen's κ coefficient.

The dataset comprised calcium oxalate monohydrate (74.9%), calcium oxalate dihydrate (17.6%), uric acid (2.7%), struvite (2.3%), and other compositions (2.5%). The CNN achieved an internal accuracy of 89.8% (AUC 0.92). In independent temporal validation the cohort accuracy was 85.4%, with an AUC of 0.81 and a Cohen's κ of 0.864. Misclassifications occurred mainly between calcium oxalate subtypes due to overlapping radiological features.

CNN-based analysis of NCCT images provides accurate, non-invasive prediction of urinary stone composition, with performance comparable to conventional methods. This approach may improve preoperative planning and broaden diagnostic access, particularly in settings where FTIR is unavailable. Prospective multicenter studies are warranted to validate clinical implementation.